Describe the bug
While reviewing the MeshNet DataLoader, specifically the SamplesDataset class, I came across a section of the code that I believe may not be functioning as intended regarding the input_length_sequence parameter. Here is a brief overview of my observations:
Observed Behavior
The SamplesDataset class is designed to handle sequences of input features with a specified input_length_sequence. However, upon examining the getitem method, it appears that the implementation assumes input_length_sequence=1 regardless of the value provided during initialization. Here is the relevant portion of the code:
def __getitem__(self, idx): ... # Prepare training data. Assume input_sequence_length=1 positions = self._data[trajectory_idx]["pos"][time_idx - 1] # (nnode, dimension) ...
Potential Issue
The comment and subsequent lines in the getitem method suggest that the code is hardcoded to handle only a single time step (input_length_sequence=1). This seems to ignore the input_length_sequence parameter provided during the initialization of the SamplesDataset class. As a result, the functionality intended to support longer input sequences may not be operational.